Abstract: There are many mountainous and plateau areas in southwest China where landslides are prone to occur frequently.It is of great significance to obtain landslide information by remote sensing technology in time for emergency rescue.Aiming at the inadequate accuracy of existing landslide detection methods and the problem of misidentification in the case of road, building and other ground objects interference, a siamese residual network landslide detection method combined with bottleneck attention was proposed.In this method, a siamese residual encoder was used to extract the feature information of remote sensing images in different periods, and a bottleneck attention module was used in the skip connection to highlight the landslide target.Finally, the spatial dimension was recovered by repeated convolution and up-sampling to complete the landslide recognition and output the recognition results.The self-made Jiuzhaigou landslide data set was used for verification.The results showed that the proposed landslide detection method had improved all the indexes compared with the original algorithms.Compared with other change detection models, this model had lower misjudgment rate and more accurate landslide detail identification results.At the same time, the Ludian earthquake landslide data set was used to test, the result showed that this method also achieved the highest recognition accuracy.In summary, the proposed method can effectively achieve high-precision and large-scale landslide hazard detection.